work

Production tools, platforms, and the occasional physical install. Each case study ends with the thing I learned that wasn't obvious going in.

platforms · 2022–presentLexus & Toyota build & priceTwo brands, one aging configurator, and an asset pipeline that topped out in the hundreds.jd power best in class · ~700k monthly usersai tools · 2026Jira velocity intelligenceJira forgets your velocity after six sprints. The fix was never a bigger context window.420k-token problem → postgres + apiai tools · 2024–presentT1 triage botEight scrum teams, a rotating triage duty, and severity calls that depended on who was on shift.~90% less manual routing · ~99% accepted · v2: inputs firstai tools · 2026compassThe AI transformation program lived in one spreadsheet, two dead apps, and a Jira project nobody owned. Compass is the program as one system — with Jira as the truth.439 items · 91% unowned → a hub with ownersplatforms · 2026beaconUsage telemetry for a fleet of internal AI tools — because a browser pixel can't see an LLM call, a cron job, or a Slack bot.18 apps reporting · 3 clients · 361 linesplatforms · 2026rotation-proofThe fleet died on schedule. Managed Postgres rotated its password weekly; the platform froze credentials at deploy; apps nobody touched went down — until credentials became a runtime concern.~1,150 restarts → one pattern, fleet-wideai tools · 2026MERIDIANThe AI-visibility prototype had grown to ~40 overlapping screens. Meridian is the clean-room successor: multi-tenant on migration one, provenance enforced by the database and CI, and a queue that is also the database.7 locked adrs · 155 commits · 23 daysai tools · 2026journey analystAn analyst you can interview about six years of lexus.com behavior — that never makes things up. Python and SQL are the calculator; the model is the translator; a second model audits every number.~6.5 years of data · 2 models · 1 ruleai tools · 2026lionbox searchLegal discovery over fifteen years of client work in enterprise Box. The AI assistant said it couldn't certify the pull was exhaustive — so completeness became a property of the crawl, and the agent was forbidden from judging it.74 ac tests · $0.11 probe within 4% · 0 contract violations in 24 runsai tools · 2026deliverables builderEighty-four broadcast deliverable codes took a producer eight unbroken hours. Eighty-four is really twenty-one decisions — so that's all anyone types now, and the acceptance test is her real sheet, cell for cell.84 → 21 decisions · ~1,940 hand-typed cells · cell-for-cell testplatforms · 2024–presentGR GT digital ecosystemA multi-year, phased ecosystem architecture for Gazoo Racing's US brand launch — built so future vehicle programs inherit it.multi-year roadmap · 12+ teams and vendorsexperiential · 2025–presentin-dealer Unreal LED wallA dealership LED video wall experience, run like a governed platform instead of a one-off activation.hardware + software + rollout roadmaporg design · 2022–presentfounding the product owner disciplineOne agency-side PO role, designed into a governed team of six.1 role → a team of 6 · 98-person org · product leads, 2026org design · 2024–presentAI governance & council frameworksMoving an agency from AI experimentation to operational capability — with guardrails, a council with a system of record, and a training series that teaches delegation, not prompting.400 trained · 150–200 in active adoption · season 2 liveai tools · 2024–presentSLA & delivery monitoringDelivery accountability was tracked manually and invisibly. Now it is automated and visible.ai tools · 2026GEOBrands get described by AI engines whether they manage it or not. GEO audits and tracks how.chatgpt · perplexity · gemini · copilotai tools · 2026AETHERCampaign intelligence: an ML pipeline does the math, an LLM writes the assessment.2 live client brands · grounded chatai tools · 2026CraggleRock + RFI acceleratorRFI assistance as a real system: retrieval, chat, gateway, and a document pipeline — seven workloads, one product.crag retrieval · 7 workloadsai tools · 2026proofing guardrailAI-assisted editorial QA where every model suggestion passes a meaning-drift gate before a human sees it.deterministic + semantic · audit receiptsai tools · 2026natural-language data lake searchAsk the media data lake a question in plain English; it finds the right view, writes the SQL, and answers.question → catalog → sql → answerai tools · 2026intake botProduct owners were the intake queue — every DM and hallway ask landed on them, and the work entered the backlog already shaped wrong.probe · capture · confirm deterministicallyplatforms · 2026pulseAd tags fail silently. Pulse watches the measurement layer itself — across lexus.com and the dealer network — and keeps the history the vendors won't.4 sources · ~160 live activities · ~244 dealer sitesai tools · 2026sprint trackerThe velocity dashboard, graduated: AI sprint reports, feature buckets, and an OKR scorecard that replaced a product owner's quarterly spreadsheet — with a fail-closed bridge to a second app.6 KRs live · 2 reproduced to the decimal · bridge fails closedplatforms · 2026release notes automationIf deployments are structured, the release notes already exist — this service writes them down.platforms · 2026program plannerTimelines, resourcing, and roadmaps with one source of truth — and a client-ready view of the same data. Then it absorbed a forecasting tool and grew a second workspace.team view + client view, one model · +forecasting · 2 workspacesplatforms · 2026blue angelsA revenue pipeline tracker — because new business deserves the same governance as delivery.ai tools · 2026trafficstation mapperBroadcast ad ops: station extraction from media buy reports — with the AI reserved for the genuinely unknown.deterministic-first · team-wide dedup cacheai tools · 2026graduating the prototypesTwo colleague-built AI-Studio apps — a programmatic-ad screenshot deck builder and a CMS component-usage auditor — rebuilt onto the platform as owned systems. Not like-for-like moves.2–3 days per deck → minutes · “never fake the page”ai tools · 2026the knowledge spineOne version-controlled source of truth for ways-of-working — serving four agent CLIs, the production tools built on it, and now a layered context model from global hub to gitignored local.1 repo → 4 agent clis + the fleet · 4 context layersplatforms · 2026the migration kitFifteen apps moved from consumer hosting to the agency's internal platform — and the process itself became the product. The first took four days; the last three took one.15 apps re-platformed · 19 numbered lessons · 1 runbookplatforms · 2026capturing a system before it goes darkA departing expert, a vehicle-data authoring service nobody else fully understood, and two Atlassian instances of scattered history — harvested, classified, and published as ten reference pages in a day.674 wiki items · ~5,760 issues · 10 reference pagesai tools · 2026fetchOffice dog-day bookings, rebuilt from a stretched-thin commerce plugin into a real system — waitlists, vaccination records, and all.83kb of acceptance criteria · 10 slots a daylab · 2026duffmate.comA four-page site for an indie Apple Watch golf app — built to look legitimate, unblock App Store review, and be citable by answer engines. Three weeks, no framework, no CMS, no analytics vendor.4 routes · 15 commits · 21 KB homepagelab · 2025NDA kioskAn unsexy compliance workflow, made into a clean system.lab · 2026synthetic c-suiteUpload the strategy, configure the executives, and watch AI personas argue about it before the real ones do — now in two editions from one image.multi-agent panel · provenance-scored · 2 editions, 1 imageai tools · 2026GEO measurement stackThree layers of AI-search measurement — and the discipline to label what each one can actually claim.clickthrough · visibility · crawlabilitydesign · 2024design system POCScalable design-token architecture, from someone who started as a designer.design · 2026design systems as agent skillsTwo design systems — the agency's and the client's — rebuilt as things an agent can invoke, not PDFs a human reads. Figma and code reconciled as two sources of truth.2 systems · 35 cards · tokens from figma + production codelab · 2011–presentpredictive market modelsTensorFlow and PyTorch models for stock and ad-buying analysis — since 2011.14 years of applied mllab · 2025–presentdelphiA personal AI stock-analysis platform — daily predictions, always-on dashboard.lab · 2026city signalLive Dallas police and fire dispatch on one map — urgency-weighted, real-time, and built on zero API keys.live sse · zero api keyslab · 2026–presentAI guitar tone engineDescribe a tone in plain English — “Mayer’s Gravity, live” — and an agent dials it into the amp sims in real time.cold start → playing in ~9 secondsai tools · 2025–2026test plan generatorAcceptance criteria in, traceable test cases out — with a quality gate that fails the run until every case passes nine explicit criteria.68 cases · 71 of 72 ac traced · 3 loops to zero failures

Several of these were built for Team One (Publicis Groupe) on the Lexus/Toyota account. Case studies describe problems and architecture without client-confidential data; screenshots and internal metrics appear as they clear.